When someone asks ChatGPT or Gemini what to buy, the chatbot doesn't search their question. It rewrites it into its own keywords, runs those searches against the live web, and builds the answer from whatever comes back. MentionBird records those keywords for every prompt you track. They're ordinary search terms — paste them into Search Console or Ahrefs and go rank for them. Start with a free report on the searches ChatGPT runs for your buyers' questions.
Between the question and the answer there's a step nobody was measuring: the chatbot picks its own keywords.
One prompt, one platform, one recorded set of searches.
The foot-types keyword is a content brief you would never have commissioned. Every prompt you track produces keywords like it — specific, unglamorous, already deciding who gets recommended. Hand them to whoever writes your content and they'll know what to do.
Some runs come back with zero keywords. The chatbot answered from memory without searching, so no page you publish can change that answer. Knowing which prompts to skip is worth as much as knowing which to chase.
When the keywords come back about something adjacent to your question, the prompt is mis-scoped. You learn that in a day from the keywords instead of a month from flat results.
Every run records the searches the chatbot ran before answering, in the order it ran them. For this prompt Gemini ran three. They aren't paraphrases or estimates — they're the strings it sent to the web, captured as it sent them.

Same prompt, one second apart, and the two chatbots asked the web six completely different questions. Not one keyword in common. Gemini went broad and added a foot-types angle. ChatGPT went narrow — it named five brands inside its own search string, then used a site: operator to pull World Athletics stack-height regulations before recommending anything.

These are real searches against the live web, recorded exactly as the chatbot ran them. "best marathon running shoes 2026" is a keyword — it has volume, competitors, and pages ranking for it right now, and your team can go after it today with the tools already on their desk.

Keywords are recorded on every run, so you can see which ones hold, which ones the chatbot swaps out, and whether ranking for one changed who got named.
Three searches each, and not one string in common. The two chatbots didn't just phrase it differently — they asked about different things.
Thirteen distinct sources between them, and only irunfar.com and solereview.com appeared in both.
Both answers named both brands. Only one made each its headline recommendation — and the keywords are why.
Illustrative example. Results vary by category and effort.
You're about to commission content off the back of these. They should hold up.
An answer report summarises what happened. A keyword list is brief.
First month $19 with code TRY19, then $79/mo. Cancel anytime.